Introduction
Medical imaging modalities like CT and MRI often produce anisotropic volumes, where high-resolution data is captured in a single plane (in-plane) while maintaining large gaps between slices (through-plane). This anisotropy saves acquisition time but introduces artifacts that hinder clinical diagnosis and downstream algorithmic analysis. While deep learning models for volumetric super-resolution (VSR) have shown promise, they are often "brittle," failing when applied to anatomical regions or scanner types different from their training data—a challenge known as the domain gap.
Methodology
The proposed Domain Adaptable Volumetric Super-Resolution (DA-VSR) framework addresses this by enabling test-time adaptation to unseen domains without requiring ground-truth labels for the new data.
Unified Architecture
DA-VSR utilizes a modular design centered around a Unified Feature Extraction (UFE) backbone. This backbone is shared across three specialized heads:
- Through-Plane Upsampler (TPU): Performs the primary super-resolution task.
- In-Plane Upsampler (IPU): Facilitates the self-supervised adaptation process.
- Slice Refinement Net (SRN): Ensures 3D consistency and removes artifacts.
Two-Stage Training and Adaptation
- Supervised Stage: The model is initially trained on a large, labeled source dataset (e.g., lung CTs) to establish a robust baseline for medical image features.
- Self-Supervised Adaptation Stage: Upon encountering a new target volume (e.g., a kidney scan), the model leverages the high-resolution in-plane slices as a self-supervision signal. By downsampling these slices and training the model to reconstruct them, the UFE backbone adapts to the specific textures and noise characteristics of the new domain. To maintain stability, the task-specific heads are frozen during this stage, forcing the backbone to learn the domain-specific features.
Experimental Setup
The model was pre-trained on the LIDC-IDRI dataset (890 lung CT scans). To evaluate its adaptability, it was tested on three out-of-domain datasets representing different organs:
- Liver: 120 volumes.
- Colon: 30 volumes.
- Kidney: 59 volumes from the KiTS19 challenge.
Results and Discussion
DA-VSR consistently outperformed state-of-the-art methods, including SAINT, SMORE, and 3D-RCAN, across all target domains.
Quantitative Performance
On the Kidney dataset at a 4x upsampling scale, DA-VSR achieved a PSNR of 38.07 dB, a significant improvement over its non-adapted baseline (37.60 dB) and the previous SOTA, SAINT (37.70 dB). This demonstrates the effectiveness of the self-supervised adaptation stage in bridging the domain gap.
Qualitative Analysis
Visual inspections revealed that DA-VSR excels at reconstructing fine anatomical details, such as the spine and small vessels, which are often blurred or lost by other methods.
While purely self-supervised approaches like SMORE tend to produce overly smooth results, DA-VSR maintains sharp, anatomically plausible structures by leveraging its pre-trained "base knowledge."
Conclusion
DA-VSR provides a flexible and efficient solution for medical image super-resolution in diverse clinical settings. By treating in-plane data as a proxy for ground truth, it enables models to adapt to new organs and scanners on-the-fly. Future work will explore cross-modality adaptation (e.g., CT to MRI) and further optimize the test-time adaptation speed.
Summary updated